AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic Labs

Summary · 5 min read instead of 1 h 25 m

Fedus and Cubuk work at Periodic Labs, which combines AI systems, simulations and high-throughput physical labs to discover new materials. Their first lab is in Menlo Park, and the team mixes chemists, physicists, experimentalists, theorists, hardware engineers and LLM researchers. Periodic's stated goal is synthesis superintelligence.

Fedus and Cubuk argue that intelligence alone cannot produce new scientific knowledge, because ideas have to be tested against reality. Periodic builds reinforcement learning environments from its own lab data, which forces models to reason under noise, missing context and small data. Simulation such as density functional theory complements the lab but cannot replace it. They say failed experiments and the full lineage of a research campaign are scarce training data, and that frontier models will still need labs to run experiments. Periodic works with industry through forward-deployed engineers, starting with semiconductors, and aims for superconductor discovery through faster synthesis.

Minute by minute

1. 0:24

   Intelligence needs experiments to find truth. Fedus says Periodic rests on the belief that no amount of rereading papers lets a model think through every experimental outcome. Progress needs conjectures tested in an iterative loop, so the company combines AI systems, physical-world simulation and a high-throughput lab. Cubuk says no team like this has existed before and compares it to Bell Labs, where theorists, experimentalists and chemists worked together.'

2. 2:44

   RL grounded in noisy physical labs. Fedus says Periodic's reinforcement learning environments and data come from its own labs, not from answers in papers or textbooks. Unlike math, lab work involves variance, aberrant measurements, unlabeled outputs and imperfect telemetry. Standard methods like mid-training and tool-using agents still apply, but sample efficiency matters more because rollouts cannot be added at will. Cubuk adds that furnaces degrade, temperatures are uneven and vibrations disturb optical instruments.'

3. 9:21

   Building RL environments around the discovery loop. Fedus lists the steps: decide what to make, work out how to synthesize it, then characterize what came out. A full experiment is too slow and noisy to serve as one rollout, so Periodic builds agents on pieces of the data. For characterization, the model is rewarded for identifying phases in X-ray diffraction data and penalized for implausible ones. Timestamped experimental records also let it build tasks the pretrained model cannot answer from memory.'

4. 13:25

   Physics is far from solved. Asked why we lack perfect simulators, Cubuk says the rest is not just chemistry, as the Dirac story suggests. High-temperature superconductivity still has no accepted theory, and density functional theory fails with strong electron correlation. He cites Phil Anderson's point that many interacting particles behave differently from the simple laws. He sees this as a reason for AI to be applied to the field.'

5. 19:44

   Mixed phases make characterization hard. Cubuk says experiments rarely give a clear yes or no. They usually yield mixed products: leftover precursors, amorphous material and unintended phases. Two phases can match the same diffraction pattern, so the system uses thermodynamics, synthesis conditions and prior knowledge to disambiguate. Electrical, magnetic and microscopy measurements add further signal, which AI can read together more easily than humans.'

6. 30:01

   Simulation and experiment complement each other. Cubuk says some quantities are easier or more accurate to get from simulation and others from experiment. He describes density functional theory as an approximation Periodic uses to estimate stability. Even 100 million DFT trials would miss microstructure and properties such as superconducting temperature, so a filter sits between calculations and lab runs. Periodic calibrates DFT against its own experiments, and Cubuk says simulations will never be enough alone.'

7. 42:31

   Dream materials and the compute-efficiency limit. Asked about dream materials, the guests name superconductors and magnets that depend less on rare earths or hard-to-source metals like cobalt. They also point to getting near the Landauer limit for the energy efficiency of computing. On quantum computing, they say people gloss over what it would simulate, since perfect crystals and DFT's limits remain.'

8. 45:57

   Putting intelligence into every instrument. Fedus says that as the lab scaled, technicians reviewing electron microscope images became a bottleneck, and a simple capture program was not useful enough. Periodic put AI on the machines with the experiment's intent, which gives richer data. Latency, cost and human patience constrain this. Fedus says full autonomy is not the goal, only a means to high-quality data, and an AI once caught a cyclic loading error across a batch of data. Cubuk says literature data is often noisier than DFT.'

9. 53:19

   Proprietary data and negative results. Fedus says Periodic uses open and closed models, and its data lets it push beyond frontier models in some tasks and use compute more efficiently. Null results matter: the literature mostly reports successes, so Cubuk says own-lab negatives allow training a classifier. A string of failures ending in success is process-engineering data. Fedus says the aim is to train on the process of doing science, not only its final output.'

10. 1:00:15

    Frontier models will still need labs. Cubuk says even much better future frontier models will have to run experiments, because scientific discovery is by definition what a model has not been trained on. Periodic builds labs so open or closed models can use them to test ideas. Fedus separates results in math and theoretical computer science from the physical world. Both say nobody will zero-shot a room-temperature superconductor.'

11. 1:05:54

    Scaling labs and lowering the noise floor. Fedus says one finite body of lab data can be reused for training by building environments from dates, subsets of instruments or whole campaigns. Cubuk describes shortcuts such as combinatorial sputtering and pooling superconductor candidates, which have mixed results. Periodic plans new types of labs and says its hardware engineering matters more as it moves from off-the-shelf instruments to its own, for example a robotic arm resting above mixed plates.'

12. 1:10:46

    A multidisciplinary team and forward-deployed engineers. Cubuk says Periodic hires across AI research, infrastructure, computation, experiments, hardware and product, all in service of synthesis superintelligence. He wants experienced researchers to work hands-on, as Bardeen did at Bell Labs. Fedus says Periodic now takes its tools to industry, with a focus on semiconductors, using forward-deployed engineers who work on site and train on partners' data.'

13. 1:17:18

    From copilots to priced outcomes. Fedus uses software engineering as an analogy: Copilot and early ChatGPT assisted engineers, and later tools like Codex automated much of the work. He expects Periodic to move from accelerating researchers to pricing outcomes as autonomy grows. Cubuk says he hopes profitable tools will draw more attention and students to solid state physics, as Bell Labs had commercial impact. They also list open-source contributions and an academic grant program.'

14. 1:20:52

    Synthesis is the bottleneck for superconductors. Cubuk says it is not hard to think of chemical spaces that could host superconductivity; synthesis is the hard part. He recounts Alex Mueller trying nickelates before cuprates worked, and later nickelates being made superconducting in thin film. He also cites a Japanese group that tried about 30,000 materials to find magnesium diboride. With perfect automation, he says, that search could run in a month, which increases the surface area for luck.'

Published Oct 8, 2026.